Fahrmeir: Discrete failure time models

Size: px

Start display at page:

Download "Fahrmeir: Discrete failure time models"

Ronald Underwood
6 years ago
Views:

1 Fahrmeir: Discrete failure time models Sonderforschungsbereich 386, Paper 91 (1997) Online unter: Projektpartner

2 Discrete failure time models Ludwig Fahrmeir, Universitat Munchen Introduction Most methods for analyzing failure time or event history data are based on time as a continuously measured variate. A basic assumption for large parts of theory is that failure times are untied, see Andersen et al. [2]. In practice, there is always some smallest time unit, so that ties can occur. A moderate number of ties, while banned in theory, can be treated by appropriate modications. If many ties occur, e.g. due to grouping in larger time units or intervals, or if time is truly discrete, then discrete survival or failure time models are more consistent with the data. Such situations arise in medical work when patients are followed up at xed intervals like months, in certain biostatistical problems, for example human fertility studies and time to pregnancy (Scheike and Jensen [19]), or in labor market studies where duration of unemployment is measured in weeks, at best, or in months. We review parametric models and outline recent nonparametric approaches. More details, in particular for parametric models, are given e.g. in Fahrmeir and Tutz [11], ch. 9, and further references cited there. Basic concepts Let time be divided into intervals [a 0 a 1 ),[a 1 a 2 ),:::,[a q;1 a q ),[a q 1). Usually a 0 = 0 is assumed, and a q denotes the nal follow up. Identifying the discrete time index t with interval [a t;1 a t ), a discrete failure time T is considered, where T = t denotes failure within interval t =[a t;1 a t ). The basic quantity characterizing T is the discrete hazard function (t) =Pr(T = tjt t) t =1 ::: q (0.1) which is the conditional probability for the risk of failure in interval t given the interval is reached. The discrete survivor function for reaching interval t

3 2 is S(t) =Pr(T t) = t;1 Y s=1 [1 ; (s)] (0.2) and the unconditional probability for failure at t is Pr(T = t) = (t)s(t). For a homogeneous population, discrete failure time data are given by (t i i ) i =1 ::: n,wheret i =min(t i C i ) is the minimum of survival time T i and censoring time C i, and i is the indicator variable for failure ( i =1) or censoring ( i =0). For the following we assume that censoring occurs at the end of the intervals, otherwise appropriate modications have tobe made. Simple estimates for (t) arecrude death rates ^(t) =d t =n t, where n t is the size of the population at risk and d t the number of observed failures in [a t;1 a t ). The so-called standard life table estimate replaces n t by n t ; w t =2, where w t is the number of censored observations in [a t;1 a t ), thereby assuming that censored observations are under risk for half the interval. In particular for large t, where the size n t of the risk set often becomes small, these estimates may be quite unsteady, and smoothing by one of the nonparametric methods outlined further below will be appropriate. This is illustrated in Figure 1. Figure 0.1: Posterior mean estimates (solid line) and pointwise two standard deviation condence bands together with crude death rates (plus) It shows crude death rates at age t in years for a population of retired American white females together with a smoothed estimate. A look at the data

4 3 (Green and Silverman [13], p. 101) shows that n t becomes rather small for higher age t, resulting in unstable estimates towards the end of the observation period. Discrete failure time data can also be described by discrete-time counting processes N i (t) i=1 ::: n, dened by N i (0) = 0 and ( 1 if individual i is at risk and fails at t N i (t) =N i (t) ; N i (t ; 1) = 0 else for t 1, see e.g. Arjas and Haara [3]. Thus, for every individual i under risk at t, the value N i (t) can be considered as the outcome P of a binary experiment, with Pr(N i (t) = 1) = (t). The sum N(t) = N i i(t) counts the number of observed failures up to t, and crude death rates can be derived as nonparametric maximum likelihood estimators, in analogy to the Nelson- Aalen estimator for continuous time. In most studies a vector of possibly time-dependent basic or derived covariates x it is observed in addition to failure times. Time-dependent components of x it are assumed to be xed within interval t. Then the hazard function for survival time T i of individual i will generally depend on covariates and is dened by i (tjx it) =Pr(T i = tjt i t x it) t =1 ::: q (0.3) where x =(x it i1 ::: x it ) denotes the history of covariates up to time t. Expressions for the survivor function (0.2) and for Pr(T = t) have to be modied accordingly. Also, the sequence of binary experiments for N i (t) t 1, will depend on x it. Unless separate analysis for homogeneous subgroups can be carried out, it is natural to describe the dependence of conditional probabilities of failure by binary regression models. Let F t; denote the history of events registered up to time t, but excluding the failure at t, and let r i (t) denote a risk indicator with r i (t) = 1 if individual i is at risk in interval t, r i (t) = 0 otherwise. Then it will be assumed that the conditional probability of failures can be expressed as Pr(N i (t) =1jF t; )=r i (t) i (tjx it)

5 4 with hazard functions linked to a time-varying predictor it by i (tjx it )=h( it) (0.4) through a suitable link function h, for example the logistic function. The predictor it is modelled parametrically or nonparametrically as a function of time t and basic or derived covariates x it. Parametric models This section deals mainly with parametric models (0.4), where the predictor has the common linear parametric form it = z 0 it (0.5) as in generalized linear models, with the design vector z it formed from basic covariates. In many applications the linear predictor is chosen as it = 0t + x 0 it x (0.6) where x is a vector of covariate eects and 0t t=1 ::: qis a time-varying baseline eect. Model (0.6) can be written in the form (0.5) by dening z 0 it = (0 ::: 1 ::: 0 x 0 ), 0 =( it 01 ::: 0q 0 ). Other predictors are discussed x further below. Dierent discrete-time failure models are determined by choice of the link function h. Most common are discrete proportional hazards and logistic models. The discrete proportional hazards model Suppose that an underlying continuous failure time obeys a proportional hazard or relativeriskmodel 0 (t) exp(x 0 it x ). If time T can only be observed as a discrete random variable, T = t denoting failure in [a t;1 a t ), this yields the discrete proportional hazards model (tjx it )=1; exp[; exp( 0t + x 0 it x )] (0.7)

6 5 with baseline eects Z at 0t =log 0 (t)dt at;1 derived from the baseline function 0 (u) (see e.g. Kalbeisch and Prentice [17]). An alternativeformulation of (0.7) is the complementary log-log model logf; log[1;(tjx 0 it )]g = 0t +x 0 it x. The parameter vector x is unchanged by the transition to the discrete version, so that the same analysis as with the proportional hazard model is possible as far as the inuence of covariates is concerned. However, x and time-varying eects 0t have nowto be estimated jointly. If the number of intervals is large, then the dimension of 01 ::: 0q may become dangerously high, often even leading to nonexistence of ML estimates. Then more parsimonious parametric forms like polynomials 0t = 0 + :::+ k t k, piecewise constant eects or regression splines with only a few cut points are preferable. Often, cubic-linear splines of the form 0t = t + 2 (t ; t c ) 2 ; + 3 (t ; t c ) 3 ;, are useful, where (t;t c ) ; =min(t;t c 0) and t c is a cut-point. The baseline eect is cubic before t c and linear after t c.such a simple spline model is more robust against few data at the end of the observation period than polynomials, and it is a smooth function, compared to piecewise constant modelling. Of course, other forms of regression splines may be considered. Also one may use the numerically more stable B-spline basis instead of the truncated power form, compare Sleeper and Harrington [20] in a continuous-time setting. By appropriate denition of the design vector, regression spline models can also be written in linear parametric from (0.5). The logistic model An alternative model is the logistic model for the discrete hazard (tjx it )= exp( 0t + x 0 it x ) 1+exp( 0t + x 0 it x ) (0.8) considered by Thompson [21] and, in slightly dierent form, by Cox[5]. For short intervals, this model becomes rather similar to the discrete proportional hazards model. An advantage of the logistic model is that the covariate eects x can be estimated semiparametrically, considering baseline eects 0t as

7 6 nuisance parameters and leaving them unspecied as in the continuous-time proportional hazards model, see Cox and Oakes [6]. Other discrete-time failure models result for other choices of h. Very exible models are obtained if the link is an element of a parametric family of link functions. Examples are the model of Aranda-Ordaz (see Fahrmeir and Tutz [11], p. 318) or the families considered by Czado [7]. Although choice of the link function is an important issue, we feel that careful modelling of the predictor is often even more essential. To simplify discussion, we consider only two covariates x and w, where x is a continuous variable like tumor size or hormone concentration and w is binary, indicating for example sex or treatment group. Models with time-varying eects are obtained by assuming it = 0t + 1 x i + 2t w i (0.9) where 2t could be the time-varying eect of a therapy, possibly decreasing with time. Alternatively, the term 2t w i may be considered as a particular form of interaction between time t and the covariate w. The function 2t may be modelled parametrically similarly as the baseline eect 0t.Amore detailed discussion of parametric time-varying eects is in Yamaguchi [24]. If the simple linear form 1 x i for the inuence of x is too restrictive, one may also try to replace it by a nonlinear smooth function 1 (x) like in generalized additive models. As in Hastie and Tibshirani [15], one may go a step further and consider varying coecient models of the form it = 0t + 1 (x i )+ 2 (x i )w i + 3t w i : (0.10) Here the smooth function 2 may be viewed as an eect of w varying over x, or it is interpreted as an interaction term between the continuous covariate x and the binary covariate w. Without further prior knowledge it will often be dicult to specify certain parametric forms for the smooth functions in (0.9), (0.10). Instead, it will be reasonable to explore patterns with nonparametric approaches outlined in the next section and to proceed then with a simpler parametric likelihood-based inference.

8 7 Likelihood inference Under appropriate conditions on censoring and covariate processes, the loglikelihood reduces to the common form known for binary regression models. Introducing the indicators ( (0 ::: 0) i =0 y i =(y i1 ::: y it )= (0 ::: 0 1) i =1 it is proportional to l() = nx t i X i=1 s=1 fy is log i (sjx is)+(1; y is ) log[1 ; i (sjx is)]g : Arjas and Haara [3] give a careful discussion of assumptions leading to l() as a (partial) log-likelihood. They will generally hold for noninformative random censoring and time-independent orexternalcovariates, but can become critical for internal covariates. In particular, the likelihood is valid in the presence of ties, by making the weak assumption that failures at t are conditionally independent given covariates and past failures. By appropriate construction of design vectors z it, the parameters can then be estimated with software for binary regression models, and other tools of likelihood inference for these models may be adopted, see Fahrmeir and Tutz [11], ch. 9. Nonparametric approaches Often, the common assumptions of linearity, additivity and time-constancy of eects are denitely violated and parametric specications of more exible models like (0.9) or (0.10) may be dicult. In this situation nonparametric approaches provide useful tools for detecting and exploring nonlinear or timedependent eects. We outline the roughness penalty approach, leading to spline-type smoothing and related Bayesian nonparametric techniques. Other methods are based on discrete kernels (e.g. Fahrmeir and Tutz [11], ch. 5, 9), or local likelihoods (Wu and Tuma [23], in a continuous-time setting, Tutz [22]). Consider models like (0.9) or (0.10) with unknown parameter vector and unknown "smooth functions" 1 2 ::: q of time or continuous covariates. The roughness penalty approach maximizes a penalized log-likelihood

9 8 criterion pl( 1 ::: p )=l( 1 ::: p ) ; px j=1 j J( j ) where J( j ) are roughness penalties and j are smoothing parameters. A simple roughness penalty for a time-varying eect jt t=1 ::: q,is J( j )= qx s=2 ( jt ; j t;1 ) 2 : (0.11) a t ; a t;1 The same form may be used for a function j (x) of some continuous covariate x. Another common penalty is J( j )= Z [ j (x) 00 ] 2 dx leading to cubic smoothing splines (see e.g. Green and Silverman [13]). Kiefer [18] proposes a discrete proportional hazards model with time-varying eects jt and penalty function (0.11). Dannegger, Klinger and Ulm [8] usethe roughness penalty approach to explore nonlinear and time-varying eects of risk factors in a breast cancer study with monthly data. Related Bayesian nonparametric approaches put smoothness priors on jt or j (x) and estimation is based on posteriors given the data. If, for example, a random walk of rst order jt = j t;1 +(a t ; a t;1 ) 1=2 v t v t N(0 1= j ) is taken as smoothness prior for the sequence f jt g, then the posterior mode or MAP estimate is identical to the penalized likelihood estimate with penalty (0.11), see Fahrmeir [9]. Full posterior analyses can be carried out with MCMC (Markov Chain Monte Carlo) techniques, see Fahrmeir and Knorr- Held [10] and, in the related context of generalized additive models, Biller and Fahrmeir [4]. Nonparametric methods are also useful for smoothing hazard functions in the absence of covariates. The smooth curve in Figure 1 is the posterior mean estimate obtained from a Bayesian MCMC approach. The corresponding cubic spline smoother is very close, see Green and Silverman [13].

10 9 Some extensions More complex discrete-time event history data Discrete failure time models can be extended similarly as continuous-time models. Often one may distinguish between several types R 2f1 ::: mg of failure or terminating events. For example, in a medical study there may be several causes of death, or in studies on unemployment duration one may consider full time and part time jobs that end the unemployment duration. The basic quantities for models with multiple modes of failure are now causespecic hazard functions ir (tjx )=Pr(T = t R = rjt t it x ) (0.12) it i.e. conditional probabilities for failure of type r in interval t. Multicategorical response models can be used for regression analysis of cause-specic hazard functions. A common candidate for unordered events is the multinomial logit model (e.g. Allison [1]). Other discrete choice models like a probit or a nested multinomial logit model (Hill et al. [16]) may also be considered. If events are ordered, ordinal response models (e.g. Fahrmeir and Tutz [11]) are appropriate. Again parametric and nonparametric approaches are possible. Penalized likelihood and Bayesian smoothing techniques with models for time-varying eects are applied to unemployment durations in Fahrmeir and Wagenpfeil [12], Fahrmeir and Knorr-Held [10]. Discrete failure time models can also be extended to general multiepisodemultistate models or, in counting process terminology, marked point processes. Hamerle [14] studies parametric regression analysis for such discrete event history data, but generally much less theoretical or applied work has been done here. Unobserved heterogeneity and frailty models The above model specications assume that individual heterogeneity can be described by observed variables. However, it is likely that not all relevant variables are included in a regression model. The conventional approach to account for neglected heterogeneity or frailty is to include individual-specic

11 10 parameters into the predictor, i.e. modifying it to h it = it + i, and to assume that the individual-specic parameters are i.i.d. random variables from a prior density function f, i.e. a normal density. Estimation can then be based on approaches for generalized mixed models with random eects, and recent MCMC methods seem particularly well suited. However, for singlespell failure time models the estimates can be very dependent on the choice of the prior specication. More experience is needed here. The problem becomes less severe with repeated events. References [1] Allison, P. D. (1982). Discrete{time methods for the analysis of event histories. In Sociological Methodology, Vol.13, S. Leinhardt, ed., Jossey{ Bass, San Francisco, pp. 61{98. [2] Andersen, P. K., Borgan, O., Gill, R. D., and Keiding, N. (1993). Statistical models based oncounting processes. Springer Verlag, New York. [3] Arjas, E., and Haara, P. (1987). A logistic regression model for hazard: Asymptotic results. Scandinavian Journal of Statistics 14, 1{18. [4] Biller, C., and Fahrmeir, L. (1997). Bayesian spline-type smoothing in generalized regression models. To appear in Computational Statistics. [5] Cox, D. R. (1972). Regression models and life tables. Journal of the Royal Statistical Society B34, 187{220. (with discussion). [6] Cox, D. R., and Oakes, D. (1984). Analysis of survival data. Chapman and Hall, London. [7] Czado, C. (1992). On link selection in generalized linear models. Springer Lecture Notes in Statistics 78, 60{65. [8] Dannegger, F., Klinger, A., and Ulm, K. (1995). Identication of prognostic factors with censored data. Discussion Paper 11, Sonderforschungsbereich 386, Ludwig-Maximilians Universitat Munchen.

12 11 [9] Fahrmeir, L. (1994). Dynamic modelling and penalized likelihood estimation for discrete time survival data. Biometrika 81, 317{330. [10] Fahrmeir, L., and Knorr-Held, L. (1997). Dynamic discrete-time duration models: estimation via Markov Chain Monte Carlo. To appear in Sociological Methodology. [11] Fahrmeir, L., and Tutz, G. (1994). Multivariate Statistical Modelling Based on Generalized Linear Models. Springer{Verlag, New York. [12] Fahrmeir, L., and Wagenpfeil, S. (1996). Smoothing hazard functions and time-varying eects in discrete duration and competing risk models. To appear in Journal of the American Statistical Association. [13] Green, P. J., andsilverman, B. W. (1994). Nonparametric Regression and Generalized Linear Models. Chapman and Hall, London. [14] Hamerle, A. (1986). Regression analysis for discrete event history or failure time data. Statistical Papers 27, 207{225. [15] Hastie, T., and Tibshirani, R. (1993). Varying{coecient models. Journal of the Royal Statistical Society B 55, 757{796. [16] Hill, D. H., Axinn, W. G., and Thornton, A. (1993). Competing hazards with shared unmeasured risk factors. In Sociological Methodology, Vol.23, P. V. Marsden, ed., Basil Blackwell, Oxford, pp. 245{277. [17] Kalbeisch, J. D., and Prentice, R. L. (1980). The Statistical Analysis of Failure Time Data. John Wiley & Sons, New York. [18] Kiefer, N. M. (1990). Econometric methods for grouped duration data. In Panel data and labor market studies, J. Hartog, G. Ridder, and J. Theeuwes, eds., Elsevier Science Publisher, pp. 97{117. [19] Scheike, T. H., and Jensen, T. K. (1997). A discrete survival model with random eects: An application to time to pregnancy. Biometrics 53, 349{360.

13 12 [20] Sleeper, L. A., and Harrington, D. P. (1990). Regression splines in the Cox model with application to covariate eects in liver disease. Journal of the American Statistical Association 85, 941{949. [21] Thompson Jr., W. A. (1977). On the treatment of grouped observations in life studies. Biometrics 33, 463{470. [22] Tutz, G. (1995). Dynamic modelling of discrete duration data: a local likelihood approach. Report 95-15, Institut fur Quantitative Methoden, Tech. Univ. Berlin. [23] Wu, L. L., and Tuma, N. B. (1991). Assessing bias and t of global and local hazard models. Sociological Methods and Research 19, 354{387. [24] Yamaguchi, K. (1993). Modeling time{varying eects of covariates in event{history analysis using statistics from the saturated hazard rate model. In Sociological Methodology, Vol.23, P. V. Marsden, ed., Basil Blackwell, Oxford, pp. 279{317.

Fahrmeir, Knorr-Held: Dynamic discrete-time duration models. (REVISED)

Fahrmeir, Knorr-Held: Dynamic discrete-time duration models. (REVISED) Sonderforschungsbereich 386, Paper 14 (1996) Online unter: http://epub.ub.uni-muenchen.de/ Projektpartner Dynamic discrete{time duration